Dissecting Hierarchical Reasoning Models: A Mechanistic Study

Do transformers, when trained on sequential reasoning traces, build internal models of the underlying task? And if so, does the structure of those internal representations mirror the structure of the domain? We train an 8-layer transformer on Sudoku solving traces and perform a mechanistic analysis of its internal computation. We establish two results. First, the model builds a substructure world model: it does not represent the board state cell by cell, as a human analyst would expect, but organizes information around the rows, columns, and boxes that Sudoku's constraints act on. Second, we identify a naked-single circuit: a small set of dedicated neurons in the final MLP layer, each individually detecting when exactly one digit remains possible for a specific cell, and reliably promoting that digit. These findings show that the geometry of an emergent world model is shaped by the constraint algebra of the domain, not its surface presentation, and that the resulting decision circuit is sparse, monosemantic, and fully interpretable. More broadly, they demonstrate that mechanistic interpretability tools can recover an end-to-end algorithmic account of how a transformer solves a combinatorial reasoning task.

Can a shared-weight recurrent Transformer develop distinct internal roles without being partitioned into separate modules? We study this in Asymmetric Input Recurrence (AIR), a minimal two-state reasoning architecture in which the same Transformer model is reused for both updates (per literature, L and H) and the only built-in difference in the update rule is that the encoded input is injected during L-updates but not H-updates. Across Sudoku-Extreme and Maze, decoded rollouts reveal a consistent split: $\zH$ behaves like a fully committed proposal state, whereas $\zL$ retains local uncertainty and shifting intermediate structure. Freeze experiments show that this split is, in practice, related to the model's state dynamics: in Sudoku, freezing $\zH$ reduces $\zL$'s content changes whereas freezing $\zL$ increases $\zH$'s, while in Maze, freezing either state increases content changes in the other state. Ablations show that to induce specialization, the shared model needs to be able to tell the two update types apart, either from input injection asymmetry or from a separate level token. Mechanistically, attention analysis shows that L-updates are consistently more local than H-updates in both Sudoku and Maze. Together, these results show that, in a two-state recurrent setting, a clear state-identity signal can induce stable, related functional roles inside a shared-parameter recurrent Transformer. Code is available at blue{https://github.com/juchengshen/air}.

Recursive reasoning models such as Hierarchical Reasoning Model (HRM) and Tiny Recursive Model (TRM) show that small, weight-shared networks can solve compute-heavy and NP puzzles by iteratively refining latent states, but their training typically relies on deep supervision and/or long unrolls that increase wall-clock cost and can bias the model toward greedy intermediate behavior. We introduce Recursive Stem Model (RSM), a recursive reasoning approach that keeps the TRM-style backbone while changing the training contract so the network learns a stable, depth-agnostic transition operator. RSM fully detaches the hidden-state history during training, treats early iterations as detached "warm-up" steps, and applies loss only at the final step. We further grow the outer recursion depth $H$ and inner compute depth $L$ independently and use a stochastic outer-transition scheme (stochastic depth over $H$) to mitigate instability when increasing depth. This yields two key capabilities: (i) $>20\times$ faster training than TRM while improving accuracy ($\approx 5\times$ reduction in error rate), and (ii) test-time scaling where inference can run for arbitrarily many refinement steps ($\sim 20,000 H_{\text{test}} \gg 20 H_{\text{train}}$), enabling additional "thinking" without retraining. On Sudoku-Extreme, RSM reaches 97.5% exact accuracy with test-time compute (within ~1 hour of training on a single A100), and on Maze-Hard ($30 \times 30$) it reaches ~80% exact accuracy in ~40 minutes using attention-based instantiation. Finally, because RSM implements an iterative settling process, convergence behavior provides a simple, architecture-native reliability signal: non-settling trajectories warn that the model has not reached a viable solution and can be a guard against hallucination, while stable fixed points can be paired with domain verifiers for practical correctness checks.